Learning-Based Task Assignment for Automated Guided Vehicles: Applying graph neural networks to optimise task assignment in an online warehouse environment

Abstract

Efficient task assignment in multi-Automated Guided Vehicle (AGV) warehouse environments is critical for optimizing industrial logistics. In collaboration with MAXAGV, this thesis evaluates the application of reinforcement learning (RL) to address this challenge. The warehouse environment is modelled as a graph, and a Graph Neural Network (GNN) policy is trained using Proximal Policy Optimization (PPO) to assign tasks to the vehicle fleet. To capture the complex topology of the facility, which is characterized by long-range spatial configurations and lock-relations that limit standard embedding methods like Node2Vec, a novel transductive node embedding scheme trained via multiple task-specific decoders is introduced. Three core GNN architectures: Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and Graph Transformers, along with their heterogeneous extensions, are evaluated and compared against conventional heuristic baselines. The empirical results demonstrate the performance trade-offs between the learning-based architectures and traditional heuristics. Furthermore, the study addresses the broader challenges of deployment, specifically the complexities of reward shaping in real-world logistics systems and the systemic barriers to integrating learning-based methods into legacy industrial infrastructures.